Simulation Dataset and Experiment Code for Physics-Informed Deep Learning RF Bone Fracture Detection

This repository accompanies the paper "Physics-Informed Deep Learning for Non-Invasive RF Bone Fracture Detection Using 2.5D Cylindrical FDTD Simulation". Contents:- Simulation dataset: 706 paired healthy/fractured electromagnetic simulations (HDF5, 243 MB) generated using a 2.5D cylindrical FDTD engine with single-pole Cole–Cole dispersive tissue modelling. Each sample contains 21-channel time-domain signals (4,096 steps), 63 hand-crafted features, and metadata (bone site, fracture type, gap width, healing stage, frequency).- Experiment pipeline: Complete Python code to reproduce all classification results, including classical baselines (SVM, Random Forest), deep learning models (1D-CNN, BiLSTM, Transformer, PI-MBTN), physics-informed loss functions, and CORAL domain adaptation for cross-site transfer. Two anatomical sites are modelled: tibia (18 mm tissue depth, 2 GHz) and femur (59 mm tissue depth, 1 GHz). The dataset supports binary classification under two labels: fracture presence and detectability. Requirements: Python 3.10+, PyTorch 2.0+, scikit-learn 1.3+, h5py 3.9+, NVIDIA GPU recommended. See README.md for full instructions.